Unvetted AI agents in bank accounts are not just an AI problem; they are also an identity problem

Banks are restricting unvetted AI agents. Here’s what that means for every enterprise and why governance, not blanket bans, is the safer path forward.

Recently, several of the largest banks in the United States have found something unsettling inside customer accounts: Activity looked human enough to pass at a casual glance, but it was not coming from a human at all. According to Axios, major US banks are detecting unauthorized third-party AI agents operating inside customer accounts. This is usually because a customer has handed the agent their own username and password, and some of those agents are attempting actions that appear independent of what the account holder asked for.

JPMorgan has already implemented tighter security measures. As CNBC reported, CEO Jamie Dimon highlighted the risk of rogue agents blocking external agents from acting as retail customers. At its core, the problem stems less from what these agents can do and more from how they get in. They arrive with no identity of their own with the same access as the customer, and the bank is unable to differentiate the agent's actions from the customer's.

While this may seem like a consumer banking story, the banks are among the first heavily regulated institutions to run into a question that every enterprise adopting agentic AI will eventually face: When an autonomous agent appears in your systems, how do you know who it is, who sent it, what it is allowed to do, and whether it is only doing what it was asked to do?

Why banks are reaching for the off switch     

When an organization cannot see or govern something, the most reliable control left is often a broad restriction, and that is where many banks have settled. Axios reports that Morgan Stanley has started blocking third-party agents, though a spokesperson said the bank hasn't seen rogue agents on its platforms. Axios also reports that JPMorgan, Bank of America, and Citi are also restricting access.

Read-only access is a reasonable stopgap, but an agent that cannot pay a bill is closer to a dashboard than an assistant. However, the banks are not opposed to AI. According to CNBC, JPMorgan already uses agents internally, restricting agents that arrive without a verifiable identity, clear scope, or an accountable owner.

If you swap a customer's bank account for your Microsoft tenant, the pattern will feel familiar to anyone who has worked in IAM: Agents are built by business teams on low-code platforms, attached to service accounts that were never properly scoped, connected to MCP servers that were not inventoried, yet owned by nobody.

In many organizations, including those in banking, that version of the story is already unfolding quietly. It shows up as a security team pausing a promising agent program because it cannot answer the audit committee, or as a compliance owner who cannot verify who approved an agent's access.

Governing agents as identities, not features  

The middle ground everyone is looking for already has a shape. It is the same discipline identity teams have applied to employees, contractors, and service accounts for two decades: Know which agents exist, give each one an identity and an accountable owner, separate the ability to see from the ability to act, put a control point between agents and the tools they use, and keep a record that compliance teams and auditors can actually work with. SecureAI, a component of AD360, is an agentic AI security and governance platform built around exactly that discipline, carried all the way through to where agents reach your tools and data.

SecureAI's MCP gateway sits between your agents and the MCP servers they call. That lets you decide which identities may use which servers and tools, at what times, and from where; whether each request is allowed; and whether the activity is logged for auditing or blocked. DLP policies can also redact or block the sensitive content you define, and reports give compliance owners a working record of agent activity.

Since none of this helps with an agent nobody knows exists, SecureAI also discovers the agents your teams have built in Copilot Studio, Microsoft Foundry, and Gemini Enterprise Agent Platform. It flags unmanaged agents, assigns each new agent an accountable owner, and keeps permissions from drifting through periodic access certification campaigns.

AI agents need governance, not blanket restrictions

The restrictions imposed by major banks highlight a challenge enterprises cannot afford to ignore: AI agents are gaining access to sensitive systems faster than organizations can establish the controls to govern them.

But restricting agents should not be the only way to manage the risks they introduce. Organizations need to know which agents operate in their environments, who is responsible for them, what they can access, and how their actions are monitored.

With SecureAI, enterprises can establish these controls without sacrificing the productivity benefits of agentic AI. The goal isn't to stop AI agents from acting, but to ensure they act within the right boundaries.